return w;
}
+static bool
+vtn_handle_non_semantic_instruction(struct vtn_builder *b, SpvOp ext_opcode,
+ const uint32_t *w, unsigned count)
+{
+ /* Do nothing. */
+ return true;
+}
+
static void
vtn_handle_extension(struct vtn_builder *b, SpvOp opcode,
const uint32_t *w, unsigned count)
val->ext_handler = vtn_handle_amd_shader_trinary_minmax_instruction;
} else if (strcmp(ext, "OpenCL.std") == 0) {
val->ext_handler = vtn_handle_opencl_instruction;
+ } else if (strstr(ext, "NonSemantic.") == ext) {
+ val->ext_handler = vtn_handle_non_semantic_instruction;
} else {
vtn_fail("Unsupported extension: %s", ext);
}
case SpvStorageClassFunction:
case SpvStorageClassWorkgroup:
case SpvStorageClassCrossWorkgroup:
+ case SpvStorageClassUniformConstant:
val->type->stride = align(glsl_get_cl_size(val->type->deref->type),
glsl_get_cl_alignment(val->type->deref->type));
break;
*before = SpvMemorySemanticsMaskNone;
*after = SpvMemorySemanticsMaskNone;
- const SpvMemorySemanticsMask order_semantics =
+ SpvMemorySemanticsMask order_semantics =
semantics & (SpvMemorySemanticsAcquireMask |
SpvMemorySemanticsReleaseMask |
SpvMemorySemanticsAcquireReleaseMask |
SpvMemorySemanticsSequentiallyConsistentMask);
+ if (util_bitcount(order_semantics) > 1) {
+ /* Old GLSLang versions incorrectly set all the ordering bits. This was
+ * fixed in c51287d744fb6e7e9ccc09f6f8451e6c64b1dad6 of glslang repo,
+ * and it is in GLSLang since revision "SPIRV99.1321" (from Jul-2016).
+ */
+ vtn_warn("Multiple memory ordering semantics specified, "
+ "assuming AcquireRelease.");
+ order_semantics = SpvMemorySemanticsAcquireReleaseMask;
+ }
+
const SpvMemorySemanticsMask av_vis_semantics =
semantics & (SpvMemorySemanticsMakeAvailableMask |
SpvMemorySemanticsMakeVisibleMask);
if (other_semantics)
vtn_warn("Ignoring unhandled memory semantics: %u\n", other_semantics);
- vtn_fail_if(util_bitcount(order_semantics) > 1,
- "Multiple memory ordering bits specified");
-
/* SequentiallyConsistent is treated as AcquireRelease. */
/* The RELEASE barrier happens BEFORE the operation, and it is usually
*after |= SpvMemorySemanticsMakeAvailableMask | storage_semantics;
}
+static void
+vtn_emit_scoped_memory_barrier(struct vtn_builder *b, SpvScope scope,
+ SpvMemorySemanticsMask semantics)
+{
+ nir_memory_semantics nir_semantics = 0;
+
+ SpvMemorySemanticsMask order_semantics =
+ semantics & (SpvMemorySemanticsAcquireMask |
+ SpvMemorySemanticsReleaseMask |
+ SpvMemorySemanticsAcquireReleaseMask |
+ SpvMemorySemanticsSequentiallyConsistentMask);
+
+ if (util_bitcount(order_semantics) > 1) {
+ /* Old GLSLang versions incorrectly set all the ordering bits. This was
+ * fixed in c51287d744fb6e7e9ccc09f6f8451e6c64b1dad6 of glslang repo,
+ * and it is in GLSLang since revision "SPIRV99.1321" (from Jul-2016).
+ */
+ vtn_warn("Multiple memory ordering semantics bits specified, "
+ "assuming AcquireRelease.");
+ order_semantics = SpvMemorySemanticsAcquireReleaseMask;
+ }
+
+ switch (order_semantics) {
+ case 0:
+ /* Not an ordering barrier. */
+ break;
+
+ case SpvMemorySemanticsAcquireMask:
+ nir_semantics = NIR_MEMORY_ACQUIRE;
+ break;
+
+ case SpvMemorySemanticsReleaseMask:
+ nir_semantics = NIR_MEMORY_RELEASE;
+ break;
+
+ case SpvMemorySemanticsSequentiallyConsistentMask:
+ /* Fall through. Treated as AcquireRelease in Vulkan. */
+ case SpvMemorySemanticsAcquireReleaseMask:
+ nir_semantics = NIR_MEMORY_ACQUIRE | NIR_MEMORY_RELEASE;
+ break;
+
+ default:
+ unreachable("Invalid memory order semantics");
+ }
+
+ if (semantics & SpvMemorySemanticsMakeAvailableMask) {
+ vtn_fail_if(!b->options->caps.vk_memory_model,
+ "To use MakeAvailable memory semantics the VulkanMemoryModel "
+ "capability must be declared.");
+ nir_semantics |= NIR_MEMORY_MAKE_AVAILABLE;
+ }
+
+ if (semantics & SpvMemorySemanticsMakeVisibleMask) {
+ vtn_fail_if(!b->options->caps.vk_memory_model,
+ "To use MakeVisible memory semantics the VulkanMemoryModel "
+ "capability must be declared.");
+ nir_semantics |= NIR_MEMORY_MAKE_VISIBLE;
+ }
+
+ /* Vulkan Environment for SPIR-V says "SubgroupMemory, CrossWorkgroupMemory,
+ * and AtomicCounterMemory are ignored".
+ */
+ semantics &= ~(SpvMemorySemanticsSubgroupMemoryMask |
+ SpvMemorySemanticsCrossWorkgroupMemoryMask |
+ SpvMemorySemanticsAtomicCounterMemoryMask);
+
+ /* TODO: Consider adding nir_var_mem_image mode to NIR so it can be used
+ * for SpvMemorySemanticsImageMemoryMask.
+ */
+
+ nir_variable_mode modes = 0;
+ if (semantics & (SpvMemorySemanticsUniformMemoryMask |
+ SpvMemorySemanticsImageMemoryMask))
+ modes |= nir_var_mem_ubo | nir_var_mem_ssbo | nir_var_uniform;
+ if (semantics & SpvMemorySemanticsWorkgroupMemoryMask)
+ modes |= nir_var_mem_shared;
+ if (semantics & SpvMemorySemanticsOutputMemoryMask) {
+ vtn_fail_if(!b->options->caps.vk_memory_model,
+ "To use Output memory semantics, the VulkanMemoryModel "
+ "capability must be declared.");
+ modes |= nir_var_shader_out;
+ }
+
+ /* No barrier to add. */
+ if (nir_semantics == 0 || modes == 0)
+ return;
+
+ nir_scope nir_scope;
+ switch (scope) {
+ case SpvScopeDevice:
+ vtn_fail_if(b->options->caps.vk_memory_model &&
+ !b->options->caps.vk_memory_model_device_scope,
+ "If the Vulkan memory model is declared and any instruction "
+ "uses Device scope, the VulkanMemoryModelDeviceScope "
+ "capability must be declared.");
+ nir_scope = NIR_SCOPE_DEVICE;
+ break;
+
+ case SpvScopeQueueFamily:
+ vtn_fail_if(!b->options->caps.vk_memory_model,
+ "To use Queue Family scope, the VulkanMemoryModel capability "
+ "must be declared.");
+ nir_scope = NIR_SCOPE_QUEUE_FAMILY;
+ break;
+
+ case SpvScopeWorkgroup:
+ nir_scope = NIR_SCOPE_WORKGROUP;
+ break;
+
+ case SpvScopeSubgroup:
+ nir_scope = NIR_SCOPE_SUBGROUP;
+ break;
+
+ case SpvScopeInvocation:
+ nir_scope = NIR_SCOPE_INVOCATION;
+ break;
+
+ default:
+ vtn_fail("Invalid memory scope");
+ }
+
+ nir_intrinsic_instr *intrin =
+ nir_intrinsic_instr_create(b->shader, nir_intrinsic_scoped_memory_barrier);
+ nir_intrinsic_set_memory_semantics(intrin, nir_semantics);
+
+ nir_intrinsic_set_memory_modes(intrin, modes);
+ nir_intrinsic_set_memory_scope(intrin, nir_scope);
+ nir_builder_instr_insert(&b->nb, &intrin->instr);
+}
+
struct vtn_ssa_value *
vtn_create_ssa_value(struct vtn_builder *b, const struct glsl_type *type)
{
return src;
}
+static uint32_t
+image_operand_arg(struct vtn_builder *b, const uint32_t *w, uint32_t count,
+ uint32_t mask_idx, SpvImageOperandsMask op)
+{
+ static const SpvImageOperandsMask ops_with_arg =
+ SpvImageOperandsBiasMask |
+ SpvImageOperandsLodMask |
+ SpvImageOperandsGradMask |
+ SpvImageOperandsConstOffsetMask |
+ SpvImageOperandsOffsetMask |
+ SpvImageOperandsConstOffsetsMask |
+ SpvImageOperandsSampleMask |
+ SpvImageOperandsMinLodMask |
+ SpvImageOperandsMakeTexelAvailableMask |
+ SpvImageOperandsMakeTexelVisibleMask;
+
+ assert(util_bitcount(op) == 1);
+ assert(w[mask_idx] & op);
+ assert(op & ops_with_arg);
+
+ uint32_t idx = util_bitcount(w[mask_idx] & (op - 1) & ops_with_arg) + 1;
+
+ /* Adjust indices for operands with two arguments. */
+ static const SpvImageOperandsMask ops_with_two_args =
+ SpvImageOperandsGradMask;
+ idx += util_bitcount(w[mask_idx] & (op - 1) & ops_with_two_args);
+
+ idx += mask_idx;
+
+ vtn_fail_if(idx + (op & ops_with_two_args ? 1 : 0) >= count,
+ "Image op claims to have %s but does not enough "
+ "following operands", spirv_imageoperands_to_string(op));
+
+ return idx;
+}
+
static void
vtn_handle_texture(struct vtn_builder *b, SpvOp opcode,
const uint32_t *w, unsigned count)
struct vtn_value *val =
vtn_push_value(b, w[2], vtn_value_type_sampled_image);
val->sampled_image = ralloc(b, struct vtn_sampled_image);
- val->sampled_image->type =
- vtn_value(b, w[1], vtn_value_type_type)->type;
val->sampled_image->image =
vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
val->sampled_image->sampler =
struct vtn_type *ret_type = vtn_value(b, w[1], vtn_value_type_type)->type;
- struct vtn_sampled_image sampled;
+ struct vtn_pointer *image = NULL, *sampler = NULL;
struct vtn_value *sampled_val = vtn_untyped_value(b, w[3]);
if (sampled_val->value_type == vtn_value_type_sampled_image) {
- sampled = *sampled_val->sampled_image;
+ image = sampled_val->sampled_image->image;
+ sampler = sampled_val->sampled_image->sampler;
} else {
vtn_assert(sampled_val->value_type == vtn_value_type_pointer);
- sampled.type = sampled_val->pointer->type;
- sampled.image = NULL;
- sampled.sampler = sampled_val->pointer;
+ image = sampled_val->pointer;
}
- const struct glsl_type *image_type = sampled.type->type;
+ nir_deref_instr *image_deref = vtn_pointer_to_deref(b, image);
+ nir_deref_instr *sampler_deref =
+ sampler ? vtn_pointer_to_deref(b, sampler) : NULL;
+
+ const struct glsl_type *image_type = sampled_val->type->type;
const enum glsl_sampler_dim sampler_dim = glsl_get_sampler_dim(image_type);
const bool is_array = glsl_sampler_type_is_array(image_type);
nir_alu_type dest_type = nir_type_invalid;
break;
case SpvOpImageFetch:
- if (glsl_get_sampler_dim(image_type) == GLSL_SAMPLER_DIM_MS) {
+ if (sampler_dim == GLSL_SAMPLER_DIM_MS) {
texop = nir_texop_txf_ms;
} else {
texop = nir_texop_txf;
nir_tex_src srcs[10]; /* 10 should be enough */
nir_tex_src *p = srcs;
- nir_deref_instr *sampler = vtn_pointer_to_deref(b, sampled.sampler);
- nir_deref_instr *texture =
- sampled.image ? vtn_pointer_to_deref(b, sampled.image) : sampler;
-
- p->src = nir_src_for_ssa(&texture->dest.ssa);
+ p->src = nir_src_for_ssa(&image_deref->dest.ssa);
p->src_type = nir_tex_src_texture_deref;
p++;
case nir_texop_txd:
case nir_texop_tg4:
case nir_texop_lod:
- /* These operations require a sampler */
- p->src = nir_src_for_ssa(&sampler->dest.ssa);
+ vtn_fail_if(sampler == NULL,
+ "%s requires an image of type OpTypeSampledImage",
+ spirv_op_to_string(opcode));
+ p->src = nir_src_for_ssa(&sampler_deref->dest.ssa);
p->src_type = nir_tex_src_sampler_deref;
p++;
break;
/* Now we need to handle some number of optional arguments */
struct vtn_value *gather_offsets = NULL;
if (idx < count) {
- uint32_t operands = w[idx++];
+ uint32_t operands = w[idx];
if (operands & SpvImageOperandsBiasMask) {
vtn_assert(texop == nir_texop_tex);
texop = nir_texop_txb;
- (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_bias);
+ uint32_t arg = image_operand_arg(b, w, count, idx,
+ SpvImageOperandsBiasMask);
+ (*p++) = vtn_tex_src(b, w[arg], nir_tex_src_bias);
}
if (operands & SpvImageOperandsLodMask) {
vtn_assert(texop == nir_texop_txl || texop == nir_texop_txf ||
texop == nir_texop_txs);
- (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_lod);
+ uint32_t arg = image_operand_arg(b, w, count, idx,
+ SpvImageOperandsLodMask);
+ (*p++) = vtn_tex_src(b, w[arg], nir_tex_src_lod);
}
if (operands & SpvImageOperandsGradMask) {
vtn_assert(texop == nir_texop_txl);
texop = nir_texop_txd;
- (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ddx);
- (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ddy);
+ uint32_t arg = image_operand_arg(b, w, count, idx,
+ SpvImageOperandsGradMask);
+ (*p++) = vtn_tex_src(b, w[arg], nir_tex_src_ddx);
+ (*p++) = vtn_tex_src(b, w[arg + 1], nir_tex_src_ddy);
}
- if (operands & SpvImageOperandsOffsetMask ||
- operands & SpvImageOperandsConstOffsetMask)
- (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_offset);
+ vtn_fail_if(util_bitcount(operands & (SpvImageOperandsConstOffsetsMask |
+ SpvImageOperandsOffsetMask |
+ SpvImageOperandsConstOffsetMask)) > 1,
+ "At most one of the ConstOffset, Offset, and ConstOffsets "
+ "image operands can be used on a given instruction.");
+
+ if (operands & SpvImageOperandsOffsetMask) {
+ uint32_t arg = image_operand_arg(b, w, count, idx,
+ SpvImageOperandsOffsetMask);
+ (*p++) = vtn_tex_src(b, w[arg], nir_tex_src_offset);
+ }
+
+ if (operands & SpvImageOperandsConstOffsetMask) {
+ uint32_t arg = image_operand_arg(b, w, count, idx,
+ SpvImageOperandsConstOffsetMask);
+ (*p++) = vtn_tex_src(b, w[arg], nir_tex_src_offset);
+ }
if (operands & SpvImageOperandsConstOffsetsMask) {
vtn_assert(texop == nir_texop_tg4);
- gather_offsets = vtn_value(b, w[idx++], vtn_value_type_constant);
+ uint32_t arg = image_operand_arg(b, w, count, idx,
+ SpvImageOperandsConstOffsetsMask);
+ gather_offsets = vtn_value(b, w[arg], vtn_value_type_constant);
}
if (operands & SpvImageOperandsSampleMask) {
vtn_assert(texop == nir_texop_txf_ms);
+ uint32_t arg = image_operand_arg(b, w, count, idx,
+ SpvImageOperandsSampleMask);
texop = nir_texop_txf_ms;
- (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ms_index);
+ (*p++) = vtn_tex_src(b, w[arg], nir_tex_src_ms_index);
}
if (operands & SpvImageOperandsMinLodMask) {
vtn_assert(texop == nir_texop_tex ||
texop == nir_texop_txb ||
texop == nir_texop_txd);
- (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_min_lod);
+ uint32_t arg = image_operand_arg(b, w, count, idx,
+ SpvImageOperandsMinLodMask);
+ (*p++) = vtn_tex_src(b, w[arg], nir_tex_src_min_lod);
}
}
- /* We should have now consumed exactly all of the arguments */
- vtn_assert(idx == count);
nir_tex_instr *instr = nir_tex_instr_create(b->shader, p - srcs);
instr->op = texop;
is_shadow && glsl_get_components(ret_type->type) == 1;
instr->component = gather_component;
- if (sampled.image && (sampled.image->access & ACCESS_NON_UNIFORM))
+ if (image && (image->access & ACCESS_NON_UNIFORM))
instr->texture_non_uniform = true;
- if (sampled.sampler && (sampled.sampler->access & ACCESS_NON_UNIFORM))
+ if (sampler && (sampler->access & ACCESS_NON_UNIFORM))
instr->sampler_non_uniform = true;
/* for non-query ops, get dest_type from sampler type */
image.sample = NULL;
break;
- case SpvOpImageRead:
+ case SpvOpImageRead: {
image.image = vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
image.coord = get_image_coord(b, w[4]);
- if (count > 5 && (w[5] & SpvImageOperandsSampleMask)) {
- vtn_assert(w[5] == SpvImageOperandsSampleMask);
- image.sample = vtn_ssa_value(b, w[6])->def;
+ const SpvImageOperandsMask operands =
+ count > 5 ? w[5] : SpvImageOperandsMaskNone;
+
+ if (operands & SpvImageOperandsSampleMask) {
+ uint32_t arg = image_operand_arg(b, w, count, 5,
+ SpvImageOperandsSampleMask);
+ image.sample = vtn_ssa_value(b, w[arg])->def;
} else {
image.sample = nir_ssa_undef(&b->nb, 1, 32);
}
+
+ if (operands & SpvImageOperandsMakeTexelVisibleMask) {
+ vtn_fail_if((operands & SpvImageOperandsNonPrivateTexelMask) == 0,
+ "MakeTexelVisible requires NonPrivateTexel to also be set.");
+ uint32_t arg = image_operand_arg(b, w, count, 5,
+ SpvImageOperandsMakeTexelVisibleMask);
+ semantics = SpvMemorySemanticsMakeVisibleMask;
+ scope = vtn_constant_uint(b, w[arg]);
+ }
+
+ /* TODO: Volatile. */
+
break;
+ }
- case SpvOpImageWrite:
+ case SpvOpImageWrite: {
image.image = vtn_value(b, w[1], vtn_value_type_pointer)->pointer;
image.coord = get_image_coord(b, w[2]);
/* texel = w[3] */
- if (count > 4 && (w[4] & SpvImageOperandsSampleMask)) {
- vtn_assert(w[4] == SpvImageOperandsSampleMask);
- image.sample = vtn_ssa_value(b, w[5])->def;
+ const SpvImageOperandsMask operands =
+ count > 4 ? w[4] : SpvImageOperandsMaskNone;
+
+ if (operands & SpvImageOperandsSampleMask) {
+ uint32_t arg = image_operand_arg(b, w, count, 4,
+ SpvImageOperandsSampleMask);
+ image.sample = vtn_ssa_value(b, w[arg])->def;
} else {
image.sample = nir_ssa_undef(&b->nb, 1, 32);
}
+
+ if (operands & SpvImageOperandsMakeTexelAvailableMask) {
+ vtn_fail_if((operands & SpvImageOperandsNonPrivateTexelMask) == 0,
+ "MakeTexelAvailable requires NonPrivateTexel to also be set.");
+ uint32_t arg = image_operand_arg(b, w, count, 4,
+ SpvImageOperandsMakeTexelAvailableMask);
+ semantics = SpvMemorySemanticsMakeAvailableMask;
+ scope = vtn_constant_uint(b, w[arg]);
+ }
+
+ /* TODO: Volatile. */
+
break;
+ }
default:
vtn_fail_with_opcode("Invalid image opcode", opcode);
vtn_emit_memory_barrier(struct vtn_builder *b, SpvScope scope,
SpvMemorySemanticsMask semantics)
{
+ if (b->options->use_scoped_memory_barrier) {
+ vtn_emit_scoped_memory_barrier(b, scope, semantics);
+ return;
+ }
+
static const SpvMemorySemanticsMask all_memory_semantics =
SpvMemorySemanticsUniformMemoryMask |
SpvMemorySemanticsWorkgroupMemoryMask |
case SpvCapabilityShaderClockKHR:
spv_check_supported(shader_clock, cap);
+ break;
+
+ case SpvCapabilityVulkanMemoryModel:
+ spv_check_supported(vk_memory_model, cap);
+ break;
+
+ case SpvCapabilityVulkanMemoryModelDeviceScope:
+ spv_check_supported(vk_memory_model_device_scope, cap);
break;
default:
case SpvAddressingModelLogical:
vtn_fail_if(b->shader->info.stage >= MESA_SHADER_STAGES,
"AddressingModelLogical only supported for shaders");
- b->shader->info.cs.ptr_size = 0;
b->physical_ptrs = false;
break;
case SpvAddressingModelPhysicalStorageBuffer64EXT:
break;
}
- vtn_assert(w[2] == SpvMemoryModelSimple ||
- w[2] == SpvMemoryModelGLSL450 ||
- w[2] == SpvMemoryModelOpenCL);
+ switch (w[2]) {
+ case SpvMemoryModelSimple:
+ case SpvMemoryModelGLSL450:
+ case SpvMemoryModelOpenCL:
+ break;
+ case SpvMemoryModelVulkan:
+ vtn_fail_if(!b->options->caps.vk_memory_model,
+ "Vulkan memory model is unsupported by this driver");
+ break;
+ default:
+ vtn_fail("Unsupported memory model: %s",
+ spirv_memorymodel_to_string(w[2]));
+ break;
+ }
break;
case SpvOpEntryPoint:
vtn_handle_decoration(b, opcode, w, count);
break;
+ case SpvOpExtInst: {
+ struct vtn_value *val = vtn_value(b, w[3], vtn_value_type_extension);
+ if (val->ext_handler == vtn_handle_non_semantic_instruction) {
+ /* NonSemantic extended instructions are acceptable in preamble. */
+ vtn_handle_non_semantic_instruction(b, w[4], w, count);
+ return true;
+ } else {
+ return false; /* End of preamble. */
+ }
+ }
+
default:
return false; /* End of preamble */
}
vtn_handle_variables(b, opcode, w, count);
break;
+ case SpvOpExtInst: {
+ struct vtn_value *val = vtn_value(b, w[3], vtn_value_type_extension);
+ /* NonSemantic extended instructions are acceptable in preamble, others
+ * will indicate the end of preamble.
+ */
+ return val->ext_handler == vtn_handle_non_semantic_instruction;
+ }
+
default:
return false; /* End of preamble */
}
break;
}
+ case SpvOpLifetimeStart:
+ case SpvOpLifetimeStop:
+ break;
+
default:
vtn_fail_with_opcode("Unhandled opcode", opcode);
}
}
/* Set shader info defaults */
- b->shader->info.gs.invocations = 1;
+ if (stage == MESA_SHADER_GEOMETRY)
+ b->shader->info.gs.invocations = 1;
/* Parse rounding mode execution modes. This has to happen earlier than
* other changes in the execution modes since they can affect, for example,
projects / mesa.git / blobdiff